Harden any language model against jailbreaks — with a defense that lives below the prompt, where an attacker can't see, override, or strip it.
Recognition
as Defense
prabodha recognizes an input's harmful signature in the model's own activations, then reinforces refusal only for genuine attacks — cutting real jailbreaks in half at zero benign over-refusal on gemma-2-2b, where a brute-force defense refuses everything. A graded, per-model library built on the J-space × Pratyabhijñā steering research.
Read the input's harmful projection in the model's activations — below what the prompt can disguise.
Pick from a tiered library: gentle→firm prompt wrappers (any model) up to the recognition-gated activation moat (open weights).
Open weights on your own GPU, or any model via BYOK/OpenRouter. Characterize per-model, then register a hardening pack to your HF.
Steering Tools
Explore, build, and evaluate steering in multiple modes
Live Steering Studio
Real-time steering with prompt, concept, amplitude, and timing control
Steering Builder
Create custom steering from concepts or contrastive examples, export as packs
Jailbreak Lab
Defensive research: test refusal steering on adversarial prompts
Alignment Lab
Test steering toward truthfulness on TruthfulQA-style questions
Arm Comparison
Compare timing arms with effect sizes and metrics
Lens Playground
Visualize and explore the workspace band activation
System Architecture
Three interlocking systems: steering pipeline (core control), research loop (dual-closure gates), and deployment topology (library → app → site → paper).
Steering Pipeline (Core)
Modules: steering/e4_cli • steering/writer.py (plan_write, capped_delta) • steering/verifier.py (readback_verdict, entropy) • contracts/trace.py (SteerTrace) • contracts/closure.py (GateReport)
Research Loop (Dual Closure)
Modules: eval/benchmarks.py (AdvBench, TruthfulQA) • eval/behavioral.py (metrics) • eval/compare.py (gate_report)